Abstract

L. Itti, C. R. Scheier, B. Khurana, C. Koch, A Simple Model of Long-Range Interactions for the Computation of Salience, In: Proc. 3rd Annual Vision Research Conference, Fort Lauderdale, FL, May 1999. (Cited by 1)

Abstract: Combining, into a single saliency map, information from multiple feature maps (each encoding for salience in different feature types such as color, intensity or orientation at different spatial scales) poses a signal-to-noise ratio problem for the detection of salient targets among distractors. For example, while an orientation pop-out would strongly appear in an orientation discontinuity map tuned to the target, localized activity resulting from the contrast between image background and target or distractors alike would also strongly appear in intensity and color contrast maps. This equally strong activity of targets and distractors in the intensity and color channels diminishes the effective salience of the target from the orientation channel. We investigated how competition for salience within each feature type may alleviate this signal-to-noise problem. We implemented a simple model of spatial competition between salient locations in the form of iterative rectified filtering by a two-dimensional ``Difference-of-Gaussians'' filter with narrow excitatory and broad inhibitory widths (2% and 25% of the image width). Based on this model, multiple locations initially eliciting comparable responses (such as in an intensity contrast map with an orientation popout stimulus) suppressed each other, while a location initially standing out (such as in an orientation contrast map with an orientation pop-out stimulus) was greatly enhanced. The competitive process hence increased target-to-distractor salience combined over all channels, both by enhancing locally stronger signals and by suppressing spatially comparable signals. This model has been applied successfully to various visual search tasks, and may provide supportive mechanistic evidence for the results of Scheier et al., in which local versus global masking differentially affects visual search.

Keywords: Bottom-up attention ; salience ; long-range interactions ; visual search

Themes: Computational Modeling, Model of Bottom-Up Saliency-Based Visual Attention